WEBVTT - The power of Granite in business 0:00:03.320 --> 0:00:06.560 Hello, Hello, Welcome to Smart Talks with IBM, a podcast 0:00:06.600 --> 0:00:12.520 from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Glapwell. This season, 0:00:12.760 --> 0:00:15.920 we're diving back into the world of artificial intelligence, but 0:00:16.000 --> 0:00:21.920 with a focus on the powerful concept of open its possibilities, implications, 0:00:21.960 --> 0:00:25.439 and misconceptions. We'll look at openness from a variety of 0:00:25.480 --> 0:00:29.240 angles and explore how the concept is already reshaping industries, 0:00:29.760 --> 0:00:33.599 ways of doing business and our very notion of what's possible. 0:00:34.400 --> 0:00:38.760 In today's episode, Jacob Goldstein sat down with Mariam Ashuri, 0:00:39.120 --> 0:00:42.360 the Director of Product Management and Head of Product for 0:00:42.440 --> 0:00:46.840 IBM's Watson x dot AI, where she spearheads the product 0:00:46.920 --> 0:00:52.320 strategy and delivery of IBM's Watson x foundation models. She 0:00:52.400 --> 0:00:55.840 is a technologist with more than fifteen years of experience 0:00:56.200 --> 0:01:01.640 developing data driven technologies. The conversation folks on how enterprises 0:01:01.880 --> 0:01:06.040 can use technology to build and deliver greater transparency in 0:01:06.120 --> 0:01:11.240 AI With Granite. Meriam explained how Grantite can be utilized 0:01:11.520 --> 0:01:16.600 to improve efficiency across various domains. She discussed how these 0:01:16.640 --> 0:01:20.840 models are being used in real world business applications, particularly 0:01:20.920 --> 0:01:25.440 in areas like customer care, where AI can help enable quick, 0:01:25.920 --> 0:01:31.679 accurate responses based on internal company data. Meriam provided a 0:01:31.720 --> 0:01:36.479 fascinating look into how enterprises have moved from mere experimentation 0:01:36.720 --> 0:01:42.480 with generative AI to actual production, navigating challenges such as 0:01:42.520 --> 0:01:47.880 increased latency, cost, and energy consumption. She highlighted how the 0:01:47.920 --> 0:01:52.480 emerging trend of smaller models customized with proprietary data can 0:01:52.520 --> 0:01:56.200 potentially deliver high performance at a fraction of the cost, 0:01:56.880 --> 0:02:02.120 marking a significant shift in how enterprises leverage AI. Whether 0:02:02.160 --> 0:02:04.960 you're an AI enthusiast, we're a business leader looking to 0:02:05.000 --> 0:02:09.720 harness the power of artificial intelligence, this episode is packed 0:02:10.240 --> 0:02:13.919 with valuable insights and forward thinking strategies. 0:02:18.240 --> 0:02:20.200 Let's just start with your background. How did you come 0:02:20.240 --> 0:02:21.400 to work at IBM. 0:02:22.080 --> 0:02:25.560 I joined IBM right after I graduated. I have an 0:02:25.560 --> 0:02:31.680 AI background, and throughout the years, I've held many roles 0:02:31.680 --> 0:02:37.160 in design, engineering, development, research, mostly focused on AI application 0:02:37.280 --> 0:02:41.400 development and design. In my current job, I'm the product 0:02:41.440 --> 0:02:45.880 owner for What's the Next DAYI, which is the IBM 0:02:45.919 --> 0:02:50.000 platform for enterprise AI. What excites me about this job, 0:02:50.040 --> 0:02:53.840 I would say, is the technology advancements over the last 0:02:53.840 --> 0:02:57.080 eighteen months in the market. We've been witnessing how generative 0:02:57.120 --> 0:02:59.720 II has been changing the market. The way that I 0:02:59.760 --> 0:03:03.160 see that is JENAI has been perhaps one of the 0:03:03.280 --> 0:03:07.079 largest paradigm shifts when we think about productivity. The same 0:03:07.120 --> 0:03:12.520 way that Internet and personal computers impacted the productivity of workforce, 0:03:12.720 --> 0:03:17.560 Now we are witnessing another wave of all those opportunities 0:03:17.600 --> 0:03:20.840 that it can unlock for especially enterprise AI when it 0:03:20.880 --> 0:03:24.960 comes to enhancing the productivity of the workforce and releasing 0:03:25.040 --> 0:03:29.680 some time that can potentially be put into creating more 0:03:29.800 --> 0:03:34.239 value work for enterprise. So that's the major part that 0:03:34.360 --> 0:03:38.200 I picked this team to have an impact on the 0:03:38.240 --> 0:03:42.280 market and the community, but also of course using the 0:03:42.880 --> 0:03:45.840 skills that I gain through all these years through IBM 0:03:46.000 --> 0:03:49.960 to help to establish IBM as the market leader for 0:03:50.120 --> 0:03:50.800 enterprise AI. 0:03:51.440 --> 0:03:55.520 So you talked about jenai as this sort of generational, 0:03:55.680 --> 0:04:00.760 transformational technological force, and I'm curious just in terms of 0:04:01.200 --> 0:04:03.640 how it's going to come into the world, Like, how 0:04:03.640 --> 0:04:07.720 do you see market adoption of genai sort of evolving 0:04:07.760 --> 0:04:08.240 from here? 0:04:09.040 --> 0:04:12.640 Well, last year was the year of excitement about generative AI. 0:04:12.760 --> 0:04:15.880 Most of the companies were experimenting and exploring with GENI. 0:04:16.600 --> 0:04:20.200 We see that energy shifted towards how to best monetize 0:04:20.200 --> 0:04:23.040 that technology. Almost half of the market has moved from 0:04:23.360 --> 0:04:28.560 investigation to pilots, ten percent has moved to production. When 0:04:28.600 --> 0:04:32.240 you're exploring with this technology, you're looking for a valve factor, 0:04:32.520 --> 0:04:36.039 You're looking for an AHA moment. That's why very large 0:04:36.080 --> 0:04:40.680 general purpose models shine. But as companies move toward production 0:04:40.760 --> 0:04:43.680 and scale, they soon realized the past success is not 0:04:43.720 --> 0:04:48.760 that straightforward. For example, they're larger the model, the larger 0:04:48.839 --> 0:04:53.400 computer resources it requires. That translates to increased latency that's 0:04:53.400 --> 0:04:57.800 your response time. That translates to increased cost. That translates 0:04:57.839 --> 0:05:01.960 to increase carbon, food print, energy consumption. So think about that. 0:05:02.120 --> 0:05:05.560 At the scale of enterprise in production, some of them 0:05:05.920 --> 0:05:09.719 can be a showstopper. Because of this reason, what actually 0:05:10.000 --> 0:05:15.200 c is emerging in the market is instead of focusing 0:05:15.240 --> 0:05:20.520 on very large general purpose models, coming back to very small, 0:05:21.520 --> 0:05:26.400 trustworthy models that they can customize on their own proprietary 0:05:26.480 --> 0:05:29.279 data that's the data about their customers, that the data 0:05:29.360 --> 0:05:33.960 about their specific domains to create something differentiated that is 0:05:34.080 --> 0:05:38.440 much smaller and delivers the performance that they want on 0:05:38.560 --> 0:05:41.159 a target use case for a fraction of the costs. 0:05:41.480 --> 0:05:45.360 Uh huh. So let's talk a little bit more specifically 0:05:45.400 --> 0:05:49.520 about what you're working on. Let's talk about Granite. First 0:05:49.520 --> 0:05:51.200 of all, tell me what is Granite. 0:05:51.800 --> 0:05:57.279 Granite is our industrial leading family of models, flagship IBM models. 0:05:58.040 --> 0:06:02.239 These are the models that we from scratch. When offered 0:06:02.240 --> 0:06:05.680 to our platform, we offer indemnification and we stand behind 0:06:05.680 --> 0:06:12.680 them today. It comes in four flavors, language, code, time series, 0:06:12.720 --> 0:06:19.240 and geospecial models. Granite language series is covering English, Spanish, German, 0:06:19.680 --> 0:06:24.479 Portuguese and Japanese. We have a combination of commercial and 0:06:24.720 --> 0:06:28.680 open source language models on Granite. For example, we recently 0:06:28.880 --> 0:06:34.000 released the Granite seven B language model, small powerful English model. 0:06:34.760 --> 0:06:38.080 On the code front, our models are state of the 0:06:38.200 --> 0:06:42.480 art models ranging from three billion to thirty four billion parameters. 0:06:43.120 --> 0:06:48.320 These are very powerful models that performs or outperforms in 0:06:48.320 --> 0:06:52.159 some cases the popular open source models in their right class. 0:06:52.200 --> 0:06:53.560 So very powerful models. 0:06:53.760 --> 0:06:56.440 So I get the idea a big picture about these models, 0:06:56.480 --> 0:06:58.200 but it would be helpful to just get a sense 0:06:58.200 --> 0:07:00.599 specifically of what they're doing. And you give me any 0:07:00.600 --> 0:07:04.920 specific examples of how these models are being used in 0:07:05.000 --> 0:07:07.119 businesses in the real world right now. 0:07:08.240 --> 0:07:11.520 Well, the top use cases for generative AI are really 0:07:11.600 --> 0:07:18.480 content generation, summarization, information extraction. Perhaps the most popular use 0:07:18.520 --> 0:07:22.200 case that we are seeing in enterprise is content grounded 0:07:22.280 --> 0:07:26.520 question and answering. So using these models as a base 0:07:26.800 --> 0:07:29.680 to connect them to a body of information let's say, 0:07:29.720 --> 0:07:34.040 their policies, their documents that is internal to the enterprise, 0:07:34.320 --> 0:07:38.520 and get the model to provide answers based on that question. 0:07:38.920 --> 0:07:42.880 One example of that is for customer agents customer care, 0:07:43.280 --> 0:07:47.840 when a customer is asking a question. Previously, the agent 0:07:47.920 --> 0:07:51.440 that responds to the customer had to answer the question 0:07:51.560 --> 0:07:54.360 and if they don't know the answer escalated to the product. 0:07:54.480 --> 0:07:57.960 Especially is keeping people on hold on the line to 0:07:58.120 --> 0:08:01.120 go figure out the answer for them and then come back. 0:08:01.160 --> 0:08:03.520 You can think of the time it takes to resolve 0:08:03.560 --> 0:08:07.240 an issue. But now we llms, we have an opportunity 0:08:07.280 --> 0:08:11.320 to automatically retrieve the information based on the internal documents 0:08:11.360 --> 0:08:14.360 of the company, formulate an answer, show it to the 0:08:14.440 --> 0:08:17.800 human agent, and then if they verify with the sources 0:08:17.840 --> 0:08:20.520 of aries coming from, they can just translate it directly 0:08:20.560 --> 0:08:24.680 to the customer. This is a very simple example of 0:08:24.720 --> 0:08:26.560 how it's impacting the customer care. 0:08:27.040 --> 0:08:31.520 So one big theme of this season is this idea 0:08:31.520 --> 0:08:34.920 of open and one of the things that's interesting to 0:08:35.040 --> 0:08:38.839 me about the work you're doing is you are using 0:08:38.920 --> 0:08:42.560 not only granted this model IBM developed, but you're also 0:08:42.720 --> 0:08:46.320 using third party models right from other places. So tell 0:08:46.360 --> 0:08:48.360 me about that work and how that is sort of 0:08:48.400 --> 0:08:52.480 fitting into your kind of real world typically enterprise GENAI work. 0:08:53.160 --> 0:08:56.559 When it comes to model strategy, our strategy is really 0:08:56.600 --> 0:09:01.120 focused on two pillars, multimodel and multidiplom It means that 0:09:01.200 --> 0:09:04.840 we don't believe one single model rules all the use cases. 0:09:05.080 --> 0:09:07.240 And I think at this point the market has also 0:09:07.320 --> 0:09:10.640 realized the enterprise markets in average today are using five 0:09:10.679 --> 0:09:15.560 to ten different models for different use cases. Oh interesting, 0:09:15.880 --> 0:09:18.160 So in our portfolio, if you look into what's on 0:09:18.240 --> 0:09:21.040 Extra DAYI today, we are offering a large sets of 0:09:21.280 --> 0:09:24.120 high performing, state of the art models coming from open 0:09:24.160 --> 0:09:28.359 source commercial models that we are bringing through our partners 0:09:28.679 --> 0:09:32.559 and also IBM developed models. In addition to all of these, 0:09:32.760 --> 0:09:35.760 we also have an option for bring your own model 0:09:36.080 --> 0:09:39.040 from outside the platform. Let's say you have a custom 0:09:39.120 --> 0:09:41.800 model that you made it yourself, you can bring it 0:09:41.840 --> 0:09:46.720 to the platform and really helping the customers to navigate 0:09:46.760 --> 0:09:50.440 through avoid range of models and pick the right model 0:09:50.679 --> 0:09:54.320 for their target use case. Throughout that we've been heavily 0:09:54.360 --> 0:09:57.560 working with our partners, and you know, this is the 0:09:57.600 --> 0:10:01.080 market that is evolving rapidly. We've been at the forefront 0:10:01.080 --> 0:10:03.240 of a spit to delivery. One example that I like 0:10:03.320 --> 0:10:08.760 to highlight is recently Metal released Lama four or five billion, 0:10:09.080 --> 0:10:11.640 such a powerful model. On the same day that it 0:10:11.800 --> 0:10:14.760 was released to the market, we made it available in 0:10:14.840 --> 0:10:17.880 our platform to our customers the same day. And not 0:10:17.960 --> 0:10:20.400 only we delivered it on the same day, we are 0:10:20.400 --> 0:10:24.880 offering competitive pricing but also for flexibility in where to deploy. 0:10:25.000 --> 0:10:27.920 So we are giving an option to enterprise to deploy 0:10:28.000 --> 0:10:32.240 these models on the platform of dat choice, either multi 0:10:32.240 --> 0:10:36.120 cloud it can be gcpaws as youre IBM cloud, or 0:10:36.160 --> 0:10:41.520 on premises. The same for mistrall Ai. Mistrall Ai recently 0:10:41.720 --> 0:10:44.680 released the model misroll large two on the same day 0:10:44.960 --> 0:10:47.960 we delivered that through the platform. That's an example of 0:10:48.000 --> 0:10:52.560 a commercial model Lama but open source, but large two 0:10:52.720 --> 0:10:56.120 is a commercial model that we made available through the platform. 0:10:56.720 --> 0:11:02.320 Great. So I want to talk about enterprise grade foundation models, 0:11:03.000 --> 0:11:05.880 just to get into it briefly, what's a foundation model. 0:11:06.360 --> 0:11:10.080 People associate foundation models with a large language model, but 0:11:10.200 --> 0:11:13.520 large language models are really a subset of foundation models. 0:11:13.679 --> 0:11:17.600 Large language models are focused on language, but foundation models 0:11:17.640 --> 0:11:21.480 can be code generators, can be focused on time series 0:11:21.520 --> 0:11:24.080 model we talked about, they can be images, it can 0:11:24.120 --> 0:11:29.239 be jewispecial models. So foundation model, as the term suggests, 0:11:29.400 --> 0:11:34.640 your foundations to create a series of subsequent models that 0:11:34.800 --> 0:11:38.719 can be customized for a downstream use case. And that's 0:11:38.760 --> 0:11:42.079 why they are calling them foundation models. LM is a 0:11:42.120 --> 0:11:44.720 good example of that as a subset for language that 0:11:44.800 --> 0:11:49.000 you can further customize on your specific data to get 0:11:49.040 --> 0:11:52.600 the model to do other works. So the core of 0:11:52.600 --> 0:11:56.120 these foundation models, they are basically trained on an ab 0:11:56.200 --> 0:12:00.280 third amount of data r data sets that of the 0:12:00.320 --> 0:12:03.560 institutions today are sourcing them from the Internet, So you 0:12:03.600 --> 0:12:06.440 can imagine what can potentially go to those models, and 0:12:06.480 --> 0:12:10.080 then it comes to the enterprise and they start using it. 0:12:10.480 --> 0:12:15.400 So for us also, when we started looking into in particular, 0:12:15.520 --> 0:12:19.600 it was triggered by customers asking us to provide client 0:12:19.720 --> 0:12:23.000 protections on these models, and we started thinking about, let's 0:12:23.000 --> 0:12:25.800 look into how the models are trained and if we 0:12:25.880 --> 0:12:29.760 are comfortable of fering client protections on the models that 0:12:29.800 --> 0:12:32.480 are available in the market. And guess what, for a 0:12:32.559 --> 0:12:36.640 majority of these models. There is absolutely no visibility into 0:12:36.720 --> 0:12:40.160 what data vent into those models, not much transparency into 0:12:40.200 --> 0:12:44.240 how the model trains, and the responsibility lies on you 0:12:44.320 --> 0:12:46.600 as the customers we start using those models. 0:12:46.640 --> 0:12:50.440 So just to be clear, that is presenting like potential risk, 0:12:50.600 --> 0:12:53.840 real potential risk to a company that is using these models. 0:12:54.120 --> 0:12:56.960 It is. It is a potential risk in particular for 0:12:57.200 --> 0:13:01.840 the customers in highly regulated industries. So what we did 0:13:01.880 --> 0:13:05.960 for Granite was when we started training these models from scratch, 0:13:06.040 --> 0:13:09.080 Basically we went to the corpus of data that was 0:13:09.120 --> 0:13:12.640 available to us. So, for example, the very first version 0:13:12.720 --> 0:13:16.920 of Granite was exposed to twenty persons of its data 0:13:16.960 --> 0:13:19.720 from finance and legal because we have a lot of 0:13:20.480 --> 0:13:25.000 financial institutions as our clients. We worked directly with our 0:13:25.080 --> 0:13:30.240 IBM research to identify detectors for harmful information like haytype 0:13:30.320 --> 0:13:31.959 use and profanity detectors. 0:13:32.520 --> 0:13:34.839 Okay, so we're talking about Granted, we're talking about this 0:13:35.080 --> 0:13:38.360 set of models IBM has developed. Let's talk about using 0:13:38.400 --> 0:13:43.200 Granite on Watson X compared to downloading open source models, 0:13:43.320 --> 0:13:44.240 Like how do those differ? 0:13:44.880 --> 0:13:48.520 By using Granite and Whatson X, you get two things. 0:13:48.920 --> 0:13:52.400 The first one is the client protection and in themification 0:13:52.559 --> 0:13:54.760 that we talked about. You get that if the model 0:13:54.800 --> 0:13:58.520 is consumed through our platform. And the second one is 0:13:58.800 --> 0:14:02.600 really the equos of platform capabilities that we are offering 0:14:02.720 --> 0:14:05.920 to help you create value on top of those data, 0:14:06.000 --> 0:14:09.960 so for example, bringing your data to customize granted for 0:14:10.040 --> 0:14:13.360 your own specific use case. But also one thing that 0:14:13.400 --> 0:14:16.880 I like to highlight in particular is the AI governance. 0:14:17.440 --> 0:14:20.120 So when you get one of these pre train models, 0:14:20.360 --> 0:14:23.520 you put it in front of your own users. Through 0:14:23.560 --> 0:14:28.200 the input and instructions that the user provides for the model, 0:14:28.680 --> 0:14:32.920 they can nodge the model to potentially create undesired behavior 0:14:33.240 --> 0:14:35.920 and change the behavior of the model. And because of 0:14:35.960 --> 0:14:40.200 this is extremely important to automatically document the lineage of 0:14:40.280 --> 0:14:44.760 who touched the model at one point, so if something happens, 0:14:44.800 --> 0:14:47.400 you can trace it back and see where it's coming from. 0:14:47.800 --> 0:14:52.600 And that's what's an extra governance is offering automatically documenting 0:14:52.720 --> 0:14:56.280 the lineage. When you use the grantite within the platform, 0:14:56.320 --> 0:14:58.640 you get all of those you can have the end 0:14:58.680 --> 0:15:01.960 to end governance, you can and have access to all 0:15:01.960 --> 0:15:05.720 these scalable deployment opportunities that is available for you, like 0:15:05.840 --> 0:15:08.720 to allow you deploy them on the platform of your 0:15:08.800 --> 0:15:12.440 choice that we talked about, either multiple cloud or on 0:15:12.520 --> 0:15:15.600 prem and it also helps you to have access to 0:15:15.600 --> 0:15:20.880 avoid range of model customizations approaches, promptuning, fine tuning, retrieval, 0:15:20.920 --> 0:15:24.440 augmented generations agents. There is a series of them available 0:15:24.760 --> 0:15:26.360 to use an apply to your model. 0:15:27.120 --> 0:15:31.600 This distinction between large language models and foundation models is 0:15:31.640 --> 0:15:36.120 eye opening. Miriam emphasized that foundation models can be tailored 0:15:36.120 --> 0:15:41.120 to specific tasks, but with that versatility comes a significant 0:15:41.200 --> 0:15:45.600 challenge the lack of transparency and how these models are trained. 0:15:46.400 --> 0:15:50.400 This composed a real risk, especially in highly regulated industries 0:15:50.720 --> 0:15:56.160 like finance. Essentially, by using Granite and watsonex together, enterprises 0:15:56.200 --> 0:15:58.960 get powerful and customizable tools. 0:16:00.120 --> 0:16:02.320 So let's talk about the future a little bit. What 0:16:02.400 --> 0:16:04.440 do you think are some of the big developments we're 0:16:04.560 --> 0:16:07.240 likely to see in the realm of AI models? 0:16:07.760 --> 0:16:12.080 Very good question. I feel like the generative AI of 0:16:12.160 --> 0:16:17.280 the past was powered by large language models. The generative 0:16:17.360 --> 0:16:21.640 AI of the future is going to reason, plan, act 0:16:22.040 --> 0:16:22.800 and reflect. 0:16:23.320 --> 0:16:26.720 Huh, and so I mean in the context of Granite 0:16:26.920 --> 0:16:30.360 in particular, like, what are we likely to see both 0:16:30.520 --> 0:16:32.400 you know, in the near term and in the sort 0:16:32.440 --> 0:16:33.680 of medium to long term. 0:16:34.320 --> 0:16:38.600 There are multiple elements to implement an agentic workflow that 0:16:38.640 --> 0:16:42.160 I just mentioned. One element of that is the LLM 0:16:42.240 --> 0:16:46.360 itself to be able to do the planning and reasoning 0:16:46.440 --> 0:16:50.800 and acting and doing something that we call tool calling. 0:16:51.200 --> 0:16:54.800 So basically, a series of tools are available to the model. 0:16:55.360 --> 0:16:58.240 You ask the model to call those and make a call. 0:16:58.400 --> 0:17:01.560 For example, we can say, hey, Granted, what is the 0:17:01.600 --> 0:17:07.320 weather like where Jacob lives. It's connect to web search API, 0:17:07.880 --> 0:17:10.639 look up your location. Then it's going to connect to 0:17:11.119 --> 0:17:15.440 weather API, calculate the weather and come back and formulate 0:17:15.480 --> 0:17:19.040 an answer and respond to that. So during this process, 0:17:19.600 --> 0:17:22.119 it first has to plan the task of how to 0:17:22.119 --> 0:17:25.000 answer that question, look into what are the tools that 0:17:25.040 --> 0:17:27.720 are available to it, and call them, and that's an 0:17:27.720 --> 0:17:30.400 ability of the model to do that. What we did 0:17:30.440 --> 0:17:34.560 with Granted was we expanded the granite capabilities to be 0:17:34.600 --> 0:17:38.240 able to do function callings. So for example, today we 0:17:38.600 --> 0:17:41.679 have an open source granted to an eb function calling 0:17:41.760 --> 0:17:44.719 that is available on hugging face to try on and 0:17:44.760 --> 0:17:47.360 you can grab the model and the model has capability 0:17:47.400 --> 0:17:50.679 to do the tool calling. I'm anticipating that in the 0:17:50.760 --> 0:17:55.000 near future the planning and reasoning and acting and reflecting 0:17:55.040 --> 0:17:58.120 capabilities of the large language models are going to continue 0:17:58.160 --> 0:17:58.639 to evolve. 0:18:00.119 --> 0:18:04.080 So thinking now from the point of view of buyers 0:18:04.119 --> 0:18:07.280 and users of AIS, really people who are listening kind 0:18:07.520 --> 0:18:14.200 from that perspective, as people are evaluating AI tools and solutions, 0:18:14.840 --> 0:18:17.720 what is the most important thing they should be thinking about? 0:18:17.800 --> 0:18:20.240 How do you think about kind of that process? 0:18:21.320 --> 0:18:24.560 I think they should always start with the area at 0:18:24.680 --> 0:18:28.760 which they think it would benefit from AI, and then 0:18:29.080 --> 0:18:33.080 within that area, look into what data they have available 0:18:33.240 --> 0:18:37.440 to potentially fit into those AI service architects do they 0:18:37.440 --> 0:18:41.000 have access to quality data? And the second question that 0:18:41.040 --> 0:18:42.919 they have to ask themselves is do I have a 0:18:42.960 --> 0:18:46.919 trusted partner that can supply what I need to be 0:18:46.960 --> 0:18:50.679 able to implement AI. That can be a collection of 0:18:50.720 --> 0:18:53.280 the foundation models that you're going to need, that can 0:18:53.320 --> 0:18:57.360 be a collection of the platform capabilities that the trusted 0:18:57.400 --> 0:19:01.240 partner can offer you to implement such a The thirting 0:19:01.560 --> 0:19:06.919 is go and evaluate the regulations. Does regulation allow you 0:19:07.000 --> 0:19:11.520 to applyoy AI to the specific area that you are 0:19:11.680 --> 0:19:15.320 investigating and you're targeting for AI. And the last part, 0:19:15.680 --> 0:19:18.919 but not least, is back to the principles of design thinking, 0:19:19.080 --> 0:19:23.840 what is the problem in that area? I'm solving with AI, 0:19:24.359 --> 0:19:27.680 and if AI is even appropriate, because we want to 0:19:27.720 --> 0:19:30.000 make sure that you use AI not just because it's 0:19:30.040 --> 0:19:32.840 a cool, hot toy in the market, but you are 0:19:33.040 --> 0:19:37.679 convinced that it can significantly enhance the user experience of 0:19:37.720 --> 0:19:40.760 your customers in that area. And once you have an 0:19:40.760 --> 0:19:43.639 answer to those all these four questions, then maybe you 0:19:43.680 --> 0:19:46.360 have a good candidates to start applying aiit. 0:19:47.320 --> 0:19:51.040 And what about from the side of project managers who 0:19:51.040 --> 0:19:54.080 are trying to just keep up with how fast things 0:19:54.119 --> 0:19:58.439 are changing, how fast innovation is happening, Like, what advice 0:19:58.440 --> 0:19:59.639 would you give those people? 0:20:00.280 --> 0:20:04.520 My advice would be focused on agility. This is a 0:20:04.560 --> 0:20:08.240 market that is evolving rapidly and the winners of the 0:20:08.320 --> 0:20:11.800 market would be those that are able to take advantage 0:20:11.800 --> 0:20:15.040 of the best the market can offer at any point 0:20:15.080 --> 0:20:18.040 of time. So in order to do that, they need 0:20:18.119 --> 0:20:26.359 to be open to experimentation, continuous learning, and to rapidly 0:20:26.680 --> 0:20:28.240 adopting the new ideas. 0:20:29.440 --> 0:20:32.919 And when you think about the future and GENAI, is 0:20:32.960 --> 0:20:36.840 there a particular, say problem that you are most excited 0:20:36.880 --> 0:20:37.399 to solve. 0:20:38.080 --> 0:20:40.879 I think that would be productivity. If you look into 0:20:41.000 --> 0:20:44.400 the stats that are out there, there are surveys that 0:20:44.680 --> 0:20:48.359 confirm that sixty to seventy percents of the time of 0:20:48.400 --> 0:20:54.359 our employees can be potentially enhanced to the productivity gains 0:20:54.359 --> 0:20:57.800 of generative I for example, I personally myself use my 0:20:57.880 --> 0:21:01.399 product for content generation a lot, so the time that 0:21:01.440 --> 0:21:06.440 it frees up can be potentially put into generating a 0:21:06.520 --> 0:21:10.720 higher value work. And because of that, I'm super excited 0:21:10.840 --> 0:21:15.280 with all the opportunities that it represents for enterprises to 0:21:15.720 --> 0:21:18.840 go and dedicate the time of their employees to higher 0:21:18.920 --> 0:21:20.000 value items. 0:21:20.240 --> 0:21:24.800 Great. Okay, a couple of Granite specific questions. So what 0:21:24.880 --> 0:21:27.639 are like the key things you want the world to 0:21:27.760 --> 0:21:29.640 know about Granite. 0:21:29.680 --> 0:21:35.680 Granite is open, trusted, and targeted. Two ways to think 0:21:35.720 --> 0:21:40.199 about openness. One open as open weights it's available for 0:21:40.280 --> 0:21:44.479 public to download, and the second one is open as 0:21:44.560 --> 0:21:49.439 in there is less restrictions on how the customers can 0:21:49.560 --> 0:21:52.640 legally use these models for a range of use cases. 0:21:52.760 --> 0:21:56.119 We have released Granite open source models on their Apache 0:21:56.320 --> 0:22:00.200 license that is enabling a large range of use cases. 0:22:00.600 --> 0:22:03.800 The second one was trusted. We talked about that like 0:22:03.880 --> 0:22:08.119 it's rooted in the trustworthy governance process that we established 0:22:08.119 --> 0:22:12.119 thereund how we are training these models and the responsibility 0:22:12.160 --> 0:22:14.640 that we take for these models. And the third one 0:22:14.680 --> 0:22:19.160 is targeted, targeted for enterprise. We talked about like exposing 0:22:19.160 --> 0:22:23.520 Granted to enterprise data or the domain specific Granted some 0:22:23.600 --> 0:22:26.960 of them like Cobalt Java Translation that is targeting to 0:22:27.160 --> 0:22:32.200 solve specific enterprise needs. And that's Granite, so open, trusted 0:22:32.280 --> 0:22:32.920 and targeted. 0:22:33.640 --> 0:22:35.439 So there are a lot of models out in the 0:22:35.480 --> 0:22:38.600 world all of a sudden, right, it's a crowded market. 0:22:39.200 --> 0:22:42.040 Where does Granite fit in that universe? What is the 0:22:42.080 --> 0:22:42.960 market for granted? 0:22:43.960 --> 0:22:47.840 We talked about the enterprise market shifting away from very 0:22:47.960 --> 0:22:53.480 large general purpose models to targeted, smaller models, and Granted 0:22:53.960 --> 0:22:58.159 is a small model that enterprise can pick up and 0:22:58.320 --> 0:23:03.159 customize on their proprietary data to create something that is 0:23:03.200 --> 0:23:07.679 differentiated for a target use case. So Granted is alsuited 0:23:07.760 --> 0:23:12.520 as a small domain specific business ready tailored for business 0:23:12.760 --> 0:23:17.159 and trained on enterprise data to solve enterprise questions. 0:23:17.560 --> 0:23:20.440 You mentioned small as one of the things that granted 0:23:20.560 --> 0:23:25.600 is why is that useful in some contexts for enterprise 0:23:25.680 --> 0:23:26.720 for businesses. 0:23:27.520 --> 0:23:32.000 The larger the model, the larger computer resources it requires, 0:23:32.680 --> 0:23:37.919 it translates to increased latency. That's your response time. It 0:23:37.960 --> 0:23:44.600 translates to increased cost and in translates to increased carbon 0:23:44.600 --> 0:23:49.119 footprint and energy consumption. So at the scale of enterprise transactions, 0:23:49.160 --> 0:23:51.520 when you move to production and you want to scale, 0:23:52.359 --> 0:23:57.560 some of these challenges can be multiple times stronger. Like 0:23:57.640 --> 0:24:00.920 costs can add up, the energy assumption can be a 0:24:01.000 --> 0:24:04.600 serious thing, and the latency is depending on the application, 0:24:05.320 --> 0:24:11.560 can be a showstopper and blocker because for longer, larger models, 0:24:11.560 --> 0:24:15.119 more powerful models, it just takes a way longer time 0:24:15.280 --> 0:24:17.159 to process and calculate the output. 0:24:17.240 --> 0:24:21.080 For you, we are going to finish up with a 0:24:21.119 --> 0:24:25.600 speed round and I want you to just answer with 0:24:25.680 --> 0:24:28.080 the first thing that comes to mind. Don't overthink this, Okay, 0:24:28.359 --> 0:24:31.280 complete this sentence. In five years, AI. 0:24:31.160 --> 0:24:33.600 Will be invisible. 0:24:33.920 --> 0:24:36.200 Ah, I like that. What do you mean by that? 0:24:36.720 --> 0:24:41.240 Today? AI is everywhere, But if you ask my kids 0:24:41.280 --> 0:24:44.600 at home, they know AI. But if you say very a, 0:24:44.680 --> 0:24:46.840 I like, how do you use AI? They don't know 0:24:46.880 --> 0:24:51.320 the answer because it's so blended in their life that 0:24:51.359 --> 0:24:55.159 they don't feel like it's something that they are using. 0:24:55.240 --> 0:24:57.520 They are getting used to that. So when I think 0:24:57.560 --> 0:25:01.800 of next generation and the years to that generation is 0:25:01.880 --> 0:25:06.400 so used to AI being part of their life that 0:25:06.440 --> 0:25:09.199 they feel like it's just there. That's one, and the 0:25:09.240 --> 0:25:12.679 second one is the simplicity of interaction with AI, that 0:25:12.760 --> 0:25:15.680 you don't feel like you're interacting with the system. It's 0:25:15.880 --> 0:25:19.159 just there. Like you talk to AI. Everything is automated. 0:25:19.280 --> 0:25:23.119 So I would say the simplicity and being blended to 0:25:23.880 --> 0:25:27.240 solve the right problems is the part that I'm referring 0:25:27.280 --> 0:25:31.399 to as invisible. Like Internet is everywhere and it's invisible. 0:25:31.680 --> 0:25:34.240 But we used to dial in, like you remember the 0:25:34.320 --> 0:25:38.720 dialing zone to connect the Internet. It's gone. Internet is 0:25:38.760 --> 0:25:40.480 completely invisible today. 0:25:40.359 --> 0:25:43.080 Right, Like we used to talk about logging on, right, 0:25:43.119 --> 0:25:46.080 and you don't log on anymore because you're always logged on. 0:25:46.760 --> 0:25:48.080